human-in-loop-workflow

Human-in-Loop Multi-Agent Workflow

A reusable collaboration protocol for small-to-medium projects: dual-agent pattern with a human coordinator. One agent implements, another reviews and signs off. The human is looped in at every review gate.

How to Use This Workflow

This skill bootstraps the protocol into any project. Bootstrap Mode may be used for either a fresh repository or an existing/forked repository.

Bootstrap

When the user asks to initialize, bootstrap, install, or adapt this workflow:

1. Check whether AGENTS.md and docs/agent-guidelines.md already exist.
2. If workflow docs are missing, create the missing target directories and fetch the protocol documents.
3. If existing project instructions are present, such as AGENTS.md, CLAUDE.md, README.md, or CONTRIBUTING.md, inspect them before changing anything. Do not overwrite them without explicit human approval.
4. For existing or forked projects, preserve upstream contribution rules and use this workflow as a coordination layer around them.
5. Leave docs/phases/README.md with no active phase unless the human has already approved a phase scope. Do not create docs/CURRENT.md until an implementation task is active; use docs/current-task-template.md when that live file is needed.
6. Report which files were created, skipped, or need human approval. Then stop.

Do not change product code, choose a stack, create implementation tasks, or start implementation during Bootstrap Mode.

The source repository is public, so no authentication is required when fetching the protocol:

Source: https://github.com/jameswei/human-in-loop-workflow (branch: main)

Fetch each file via its raw URL and write to the target path:

  Raw URL                                                     Target path
  ────────                                                    ───────────
  .../main/AGENTS.md               →  AGENTS.md               (project root)
  .../main/docs/agent-guidelines.md     →  docs/agent-guidelines.md
  .../main/docs/current-task-template.md  →  docs/current-task-template.md
  .../main/docs/phase-spec-template.md    →  docs/phase-spec-template.md
  .../main/docs/taskboard-template.md     →  docs/taskboard-template.md
  .../main/docs/phases/README.md          →  docs/phases/README.md

Use the simplest available method: curl, wget, the agent's native HTTP fetch tool, or git clone with sparse-checkout. After fetching, verify every file exists at its target path.

After Bootstrap

Open AGENTS.md first. It routes you to the correct reading order and the active work context. Do not re-fetch unless documents are missing or the user asks.

If no phase is active, stay in Planning Mode or Phase Setup. Do not begin implementation until the human approves a phase spec and taskboard.

Core Rules

• Follow docs/agent-guidelines.md.
• Never mark your own work as reviewed or done.
• Update docs/CURRENT.md for active implementation tasks.
• Flag unclear scope, architecture, or upstream-process conflicts for human decision.